Terry: Skills de Engenharia de Software
drubinstein/terry
Metodologias disciplinadas de engenharia como skills reutilizáveis para Claude Code
Ver no GitHubColeção de skills reutilizáveis para Claude Code que automatizam metodologias disciplinadas de engenharia de software: reprodução de bugs (jamie), debug por eliminação (sherlock), análise de performance (brendan), iteração empírica (andrej), orquestração de fleets (terry) e experimentos A/B (ronald). Cada skill é um playbook autocontido que Claude carrega sob demanda para resistir a atalhos e garantir rigor.
Instalação
npx convez-ai drubinstein/terry1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Reproduzir e confirmar bugs antes de corrigir
- Debugar falhas resistentes com análise sistemática de causa raiz
- Identificar bottlenecks de performance usando método USE
- Iterar em soluções gated a métricas objetivas
- Orquestrar múltiplas tarefas longas em paralelo
- Executar experimentos controlados A/B entre modelos ou prompts
Veredito

Skill madura e bem-documentada que codifica disciplina em tarefas de engenharia onde Claude tende a ceder a atalhos. Excelente para times que precisam repetibilidade e evidência em debugging, otimização e experimentação — não é um experimento, é um padrão consolidado.
README
A tradução para o português deste README ainda está na fila.
Claude Code skills for software engineering
A small collection of Claude Code skills — reusable, named methodologies for
everyday engineering work. Each is a self-contained playbook (a SKILL.md plus
optional references and templates) that Claude loads on demand when your request
matches it.
Why this exists
Claude writes code well but, under pressure, tends to guess — skip reproducing a bug, theorize before measuring, shrug a failure off as "flaky," or call something done without evidence. Each skill here encodes a disciplined methodology that resists exactly that, distilled from real long-running agent work and generalized. You invoke a skill by intent ("does this bug actually reproduce?", "find the bottleneck"); Claude loads the matching playbook and follows it.
The skills
| Skill | Reach for it when… |
|---|---|
| jamie | you have a bug report and need to confirm it actually reproduces before fixing it |
| sherlock | a bug resists the obvious explanation — find the cause by elimination |
| brendan | something is slow — find the bottleneck (Brendan Gregg's USE method) |
| andrej | you're iterating toward a fix or metric and don't want to guess |
| terry | you have many independent long-running tasks to run in parallel and merge |
| ronald | you're choosing between models or prompts — run a controlled A/B experiment, not a vibe check |
Most stand alone, but five compose into a debugging-to-fix pipeline: jamie (confirm
it's real) → sherlock (find the cause) → brendan (find the bottleneck) →
andrej (drive the fix, gated on a metric) → terry (run many at once). ronald
is the odd one out — a controlled experiment harness for picking a model or prompt with
evidence. See each skill's SKILL.md for the full method.
Example
You paste a bug report: "CSV export corrupts my file every time." Claude recognizes a triage task and loads jamie — which stops it from closing the issue after one green run on the latest build. Instead it matches the reporter's version and file, reproduces the crash at their data scale, and minimizes it to a one-command repro: CONFIRMED. It hands that repro to sherlock to eliminate suspects down to the root cause, then to andrej to drive the fix, gated on the repro until it passes.
Using these skills
As a plugin — point your Claude Code plugin config at this repo (or a marketplace
entry that references it). The manifest at .claude-plugin/plugin.json auto-discovers
every skills/*/SKILL.md.
As personal skills — copy any skill directory into your user skills folder:
cp -R skills/sherlock ~/.claude/skills/sherlock # or any other skill
Either way, Claude auto-discovers them and loads the right one when your request matches its description — or invoke one explicitly with the Skill tool.
Layout
Each skill is a directory under skills/:
skills/<name>/
SKILL.md # the methodology — what it's for and how to run it
references/ # deeper detail, loaded on demand (optional)
templates/ # copy-paste worksheets / scripts (optional)
The repo is named terry for historical reasons — it began as that single
agent-fleet skill — but it's now a general software-engineering skills collection,
and terry is just one of them.
Contributing
Each skill follows the same shape — a tight SKILL.md with depth in references/
and copy-paste assets in templates/ — and is built test-first: write a failing
scenario for a subagent, write the skill, then close the loopholes it finds. New
skills and fixes are welcome via pull request or
issues.
License
MIT © David Rubinstein